Cin TI I used to avoid too many negative prompts
C(if that is the context you are talking about now heheh)
Fyep
Cyes, prompts like "ugly" can make the face another one
Cmaybe it is because TI is about tokens, prompts
Cso one can change the whole thing
CI dunno really
FTI affected more than dreambooth that is for sure
Fi am gonna test just like this now
Cok
Cit is quite delicated the thing of TI
Ci have found that if you train it in the vanilla, it works better in the vanilla. But not too much in protogen. And if you train it in protogen, it works better in protogen. But not really amazingly. Just an 60%/70% ... it is quite hard... I need like 10 or 20 more tests... trial and error. Maybe you, the community can help
HThe Clip interrogator V2 is helpful for prompting
HIt designed to output good prompts for v2 based on an image
Cyeah, it is lush!
Cdelta.safetensors is an extension of a file the newest amazing extension generates... (any knows for what is that?) I think it is an useless file
Fif only face was matching i will generate some to see :d
F
F
Cwith matching what do you mean? the images you used for training were only faces?
Flol
Chahah
Fi mean generating an image that has my face and riding a dinasaour
Cnot baad, but too much cgi, 3d, 90s style
Cand... it is you at 80%?
F
Fvery similar this one
Cyeah
Flol
Cwtf
Cxd
Falmost there haah
Ctry more steps with same seed
Fdecent one
Cthat one was cool
Cso the closer the better accuracy
Clike always
Csd 2.1. TI
Cso TI works in sd 2.1
F
Fyes ti sd 2.1
Fnow will test step count on this
Ctoo much like a videogame of 1995
Fdo you know format for that
Ctry negative prompts like CGI or 3D
Fok 1-50 worked
CI find Euler makes it
Chrfhfhrf
Cand DPM2 a Karras for some strange reason makes it more accuracy
C*accurated